Variable tracking: Cleaned up and rebased.

This is the much-famed "variable tracking" patch. Rebased against current bitbake, fixed a couple of things.

WHAT IT DOES:

        $ bitbake -e > variable_history

        $ cat variable_history
        #
        # INCLUDE HISTORY:
        #
        # /home/seebs/oe-core/build/conf/bblayers.conf
        # /home/seebs/oe-core/meta/conf/layer.conf
        # conf/bitbake.conf includes:
        #   /home/seebs/oe-core/meta/conf/abi_version.conf
        #   conf/site.conf
        #   conf/auto.conf
        #   /home/seebs/oe-core/build/conf/local.conf
        [...]
        # $prefix [2 operations]
        #   exported conf/bitbake.conf:17
        #     [export] "1"
        #   set conf/bitbake.conf:17
        #     "/usr"
        # computed:
        #   "/usr"
        export prefix="/usr"
        [...]

This can get pretty useful if you're trying to find out, say, where the -m32 came from in TUNE_CCARGS:

        # $TUNE_CCARGS [6 operations]
        #   set conf/bitbake.conf:105
        #     [defaultval] ""
        #   set conf/bitbake.conf:106
        #     [vardepvalue] "${TUNE_CCARGS}"
        #   append /home/seebs/oe-core/meta/conf/machine/include/ia32/arch-ia32.inc:16
        #     "${@bb.utils.contains("TUNE_FEATURES", "m32", "-m32", "", d)}"
        #   append /home/seebs/oe-core/meta/conf/machine/include/ia32/arch-ia32.inc:24
        #     "${@bb.utils.contains("TUNE_FEATURES", "mx32", "-mx32", "", d)}"
        #   append /home/seebs/oe-core/meta/conf/machine/include/ia32/arch-ia32.inc:32
        #     "${@bb.utils.contains("TUNE_FEATURES", "m64", "-m64", "", d)}"
        #   append /home/seebs/oe-core/meta/conf/machine/include/tune-i586.inc:7
        #     "${@bb.utils.contains("TUNE_FEATURES", "i586", "-march=i586", "", d)}"
        # computed:
        #   " ${@bb.utils.contains("TUNE_FEATURES", "m32", "-m32", "", d)} ${@bb.utils.contains("TUNE_FEATURES", "mx32", "-mx32", "", d)} ${@bb.utils.contains("TUNE_FEATURES", "m64", "-m64", "", d)} ${@bb.utils.contains("TUNE_FEATURES", "i586", "-march=i586", "", d)}"
        TUNE_CCARGS="-m32   -march=i586"
        #

The cleanup since the last version is pretty minor, but there was one real
bug: If any part of a variable's history showed a modification from Python code (with a function name) rather than from the parser, the variable's value would be displayed as though it were a function. Whoops.

The output for renames is now much more useful than it was before, though:

# $SECTION_${PN}-dbg [2 operations]
# set conf/bitbake.conf:316
# "devel"
# rename (to) data.py:161 [expandKeys]
# "SECTION_core-image-minimal-dbg"
# computed:
# "None"

and

# $SECTION_core-image-minimal-dbg
# rename from SECTION_${PN}-dbg data.py:161 [expandKeys]
# "devel"
SECTION_core-image-minimal-dbg="devel"

I remain a little unhappy with the loginfo stuff, in particular the possible clash between keys in loginfo and non-keyword arguments, but I can't really think of a way to make it definitely better, and the much-less-intrusive usage strikes me as a big enough improvement to justify it. (Picking different names for the keyword arguments would reduce clashes, and also legibility in cases where they're being specified.)

The following changes since commit dee7decf87dfa8cb966fe40846d27f3e6ab1846b:
Richard Purdie (1):
build.py: Fix traceback when there are no dependencies

are available in the git repository at:

git://git.yoctoproject.org/poky-contrib seebs/incvar
http://git.yoctoproject.org/cgit.cgi/poky-contrib/log/?h=seebs/incvar

Peter Seebach (2):
data_smart.py and friends: Track file inclusions for bitbake -e data_smart.py and friends: Track variable history

 lib/bb/cooker.py                   |   15 ++-
 lib/bb/data.py                     |   38 +++++-
 lib/bb/data_smart.py               |  250 ++++++++++++++++++++++++++++++++----
 lib/bb/parse/__init__.py           |    3 +-
 lib/bb/parse/ast.py                |   25 +++-
 lib/bb/parse/parse_py/BBHandler.py |    6 +-
 6 files changed, 299 insertions(+), 38 deletions(-)

This code adds inclusion history to bitbake -e output, showing which files were included, in what order. This doesn't completely resolve timing questions, because it doesn't show you which lines of a file were processed before or after a given include, but it does let you figure out what the path was by which a particular file ended up in your build at all.

How it works: data_smart acquires a .history member, which is an IncludeHistory; this represents the inclusion of a file and all its inclusions, recursively. It provides methods for including files, for finishing inclusion (done as an __exit__), and for dumping the whole tree.

The parser is modified to run includes inside a with() to push and pop the include filename.

---
 lib/bb/cooker.py         |    5 +++++
 lib/bb/data_smart.py     |   44 ++++++++++++++++++++++++++++++++++++++++++++
 lib/bb/parse/__init__.py |    3 ++-
 3 files changed, 51 insertions(+), 1 deletions(-)

index 1d38164..86e3ed6 100644
@@ -338,6 +338,11 @@ class BBCooker:
                 parselog.exception("Unable to read %s", fn)
                 raise
 
+        # Display history
+        with closing(StringIO()) as env:
+            self.configuration.data.history.emit(env)
+            logger.plain(env.getvalue())
+
         # emit variables and shell functions
         data.update_data(envdata)
         with closing(StringIO()) as env:

index 4293d09..ae0d9f6 100644
@@ -87,7 +87,8 @@ def handle(fn, data, include = 0):
     """Call the handler that is appropriate for this file"""
     for h in handlers:
         if h['supports'](fn, data):
-            return h['handle'](fn, data, include)
+            with data.history.include(fn):
+                return h['handle'](fn, data, include)
     raise ParseError("not a BitBake file", fn)
 
 def init(fn, data):

1.7.0.4

This patch adds tracking of the history of variable assignments. The changes are predominantly localized to data_smart.py and parse/ast.py. BBHandler.py is changed to use d.setVar(...) instead of data.setVar(..., d), and cooker.py and data.py are altered to display the recorded data, and turn tracking on for the bitbake -e case. The data.py update_data() function warns DataSmart.finalize() to report the caller one further back up the tree.

In general, d.setVar() does what it used to do. Optionally, arguments describing an operation may be appended; if none are present, the operation is implicitly ignored. If it's not ignored, it will attempt to infer missing information (name of variable, value assigned, file and line) by examining the traceback. This slightly elaborate process eliminates a category of problems in which the 'var' member of the keyword arguments dict is set, and a positional argument corresponding to 'var' is also set. It also makes calling much simpler for the common cases.

If you make manual calls to some of these functions, you can get surprises; for instance, infer_caller_details will set the 'value' member of the loginfo object to None if it can't find a value; if you then pass that loginfo object to setVar, after a value, the loginfo array clashes with the positional parameter named value.

The resulting output gives you a pretty good picture of what values got set, and how they got set.

---
 lib/bb/cooker.py                   |   10 ++-
 lib/bb/data.py                     |   38 ++++++-
 lib/bb/data_smart.py               |  216 +++++++++++++++++++++++++++++++-----
 lib/bb/parse/ast.py                |   25 ++++-
 lib/bb/parse/parse_py/BBHandler.py |    6 +-
 5 files changed, 253 insertions(+), 42 deletions(-)

index 86e3ed6..35e79a7 100644
@@ -101,8 +101,10 @@ class BBCooker:
         # to use environment variables which have been cleaned from the
         # BitBake processes env
         self.savedenv = bb.data.init()
+        if self.configuration.show_environment:
+            self.savedenv.enableTracking()
         for k in savedenv:
-            self.savedenv.setVar(k, savedenv[k])
+            self.savedenv.setVar(k, savedenv[k], op = 'inherit', file = '[saved environment]')
 
         self.caches_array = []
         # Currently, only Image Creator hob ui needs extra cache.

@@ -177,6 +179,8 @@ class BBCooker:
 
     def initConfigurationData(self):
         self.configuration.data = bb.data.init()
+        if self.configuration.show_environment:
+            self.configuration.data.enableTracking()
 
         if not self.server_registration_cb:
             self.configuration.data.setVar("BB_WORKERCONTEXT", "1")

@@ -186,6 +190,8 @@ class BBCooker:
 
     def loadConfigurationData(self):
         self.configuration.data = bb.data.init()
+        if self.configuration.show_environment:
+            self.configuration.data.enableTracking()
 
         if not self.server_registration_cb:
             self.configuration.data.setVar("BB_WORKERCONTEXT", "1")

@@ -846,6 +852,8 @@ class BBCooker:
 
     def parseConfigurationFiles(self, prefiles, postfiles):
         data = self.configuration.data
+        if self.configuration.show_environment:
+            data.enableTracking()
         bb.parse.init_parser(data)
 
         # Parse files for loading *before* bitbake.conf and any includes

index 478a482..ec0db9e 100644
@@ -166,9 +166,9 @@ def inheritFromOS(d, savedenv, permitted):
     for s in savedenv.keys():
         if s in permitted:
             try:
-                setVar(s, getVar(s, savedenv, True), d)
+                d.setVar(s, getVar(s, savedenv, True), op = 'inherit')
                 if s in exportlist:
-                    setVarFlag(s, "export", True, d)
+                    d.setVarFlag(s, "export", True, op = 'automatic')
             except TypeError:
                 pass
 

@@ -194,8 +194,38 @@ def emit_var(var, o=sys.__stdout__, d = init(), all=False):
         return 0
 
     if all:
+        history = d.history.variable(var)
         commentVal = re.sub('\n', '\n#', str(oval))
-        o.write('# %s=%s\n' % (var, commentVal))
+        if history:
+            if len(history) == 1:
+                o.write("#\n# $%s\n" % var)
+            else:
+                o.write("#\n# $%s [%d operations]\n" % (var, len(history)))
+            for event in history:
+                # o.write("# %s\n" % str(event))
+                if 'details' in event and event['details']:
+                    value = event['details']
+                else:
+                    value = event['value']
+                if 'func' in event:
+                    # If we have a function listed, this is internal
+                    # code, not an operation in a config file, and the
+                    # full path is distracting.
+                    event['file'] = re.sub('.*/', '', event['file'])
+                    display_func = ' [%s]' % event['func']
+                else:
+                    display_func = ''
+                if 'flag' in event:
+                    flag = '[%s] ' % (event['flag'])
+                else:
+                    flag = ''
+                o.write("#   %s %s:%s%s\n#     %s\"%s\"\n" % (event['op'], event['file'], event['line'], display_func, flag, re.sub('\n', '\n#     ', value)))
+            if len(history) > 1:
+                o.write("# computed:\n")
+                o.write('#   "%s"\n' % (commentVal))
+        else:
+            o.write("#\n# $%s\n#   [no history recorded]\n#\n" % var)
+            o.write('#   "%s"\n' % (commentVal))
 
     if (var.find("-") != -1 or var.find(".") != -1 or var.find('{') != -1 or var.find('}') != -1 or var.find('+') != -1) and not all:
         return 0

@@ -274,7 +304,7 @@ def emit_func(func, o=sys.__stdout__, d = init()):
 
 def update_data(d):
     """Performs final steps upon the datastore, including application of overrides"""
-    d.finalize()
+    d.finalize(parent = True)
 
 def build_dependencies(key, keys, shelldeps, vardepvals, d):
     deps = set()

index 79e9f7e..ba672b1 100644
@@ -68,7 +68,7 @@ class ExportNode(AstNode):
         self.var = var
 
     def eval(self, data):
-        data.setVarFlag(self.var, "export", 1)
+        data.setVarFlag(self.var, "export", 1, op = 'exported')
 
 class DataNode(AstNode):
     """

@@ -90,33 +90,48 @@ class DataNode(AstNode):
     def eval(self, data):
         groupd = self.groupd
         key = groupd["var"]
+        loginfo = {
+            'variable': key,
+            'file': self.filename,
+            'line': self.lineno,
+        }
         if "exp" in groupd and groupd["exp"] != None:
-            data.setVarFlag(key, "export", 1)
+            data.setVarFlag(key, "export", 1, op = 'exported', **loginfo)
+        # The others all want this by default:
+        loginfo['details'] = groupd["value"]
+        op = None
         if "ques" in groupd and groupd["ques"] != None:
             val = self.getFunc(key, data)
+            op = "set?"
             if val == None:
                 val = groupd["value"]
         elif "colon" in groupd and groupd["colon"] != None:
             e = data.createCopy()
             bb.data.update_data(e)
+            op = "immediate"
             val = e.expand(groupd["value"], key + "[:=]")
         elif "append" in groupd and groupd["append"] != None:
+            op = "append"
             val = "%s %s" % ((self.getFunc(key, data) or ""), groupd["value"])
         elif "prepend" in groupd and groupd["prepend"] != None:
+            op = "prepend"
             val = "%s %s" % (groupd["value"], (self.getFunc(key, data) or ""))
         elif "postdot" in groupd and groupd["postdot"] != None:
+            op = "postdot"
             val = "%s%s" % ((self.getFunc(key, data) or ""), groupd["value"])
         elif "predot" in groupd and groupd["predot"] != None:
+            op = "predot"
             val = "%s%s" % (groupd["value"], (self.getFunc(key, data) or ""))
         else:
             val = groupd["value"]
 
+        loginfo['op'] = op
         if 'flag' in groupd and groupd['flag'] != None:
-            data.setVarFlag(key, groupd['flag'], val)
+            data.setVarFlag(key, groupd['flag'], val, **loginfo)
         elif groupd["lazyques"]:
-            data.setVarFlag(key, "defaultval", val)
+            data.setVarFlag(key, "defaultval", val, **loginfo)
         else:
-            data.setVar(key, val)
+            data.setVar(key, val, **loginfo)
 
 class MethodNode(AstNode):
     def __init__(self, filename, lineno, func_name, body):

index 2ee8ebd..c585b60 100644
@@ -84,7 +84,7 @@ def inherit(files, fn, lineno, d):
         if not file in __inherit_cache:
             logger.log(logging.DEBUG -1, "BB %s:%d: inheriting %s", fn, lineno, file)
             __inherit_cache.append( file )
-            data.setVar('__inherit_cache', __inherit_cache, d)
+            d.setVar('__inherit_cache', __inherit_cache)
             include(fn, file, lineno, d, "inherit")
             __inherit_cache = d.getVar('__inherit_cache') or []
 

@@ -134,7 +134,7 @@ def handle(fn, d, include):
         __inherit_cache = d.getVar('__inherit_cache') or []
         if not fn in __inherit_cache:
             __inherit_cache.append(fn)
-            data.setVar('__inherit_cache', __inherit_cache, d)
+            d.setVar('__inherit_cache', __inherit_cache)
 
     if include != 0:
         oldfile = d.getVar('FILE')

@@ -151,7 +151,7 @@ def handle(fn, d, include):
 
     # DONE WITH PARSING... time to evaluate
     if ext != ".bbclass":
-        data.setVar('FILE', abs_fn, d)
+        d.setVar('FILE', abs_fn)
 
     try:
         statements.eval(d)

1.7.0.4

I *really* want to take these, however I did want to check performance:

Before:

$ time bitbake core-image-sato -e
real	0m8.034s
$ time bitbake -p
real	0m12.220s
user	3m3.887s

After:

$ time bitbake core-image-sato -e
real	0m50.267s
$ time bitbake -p
real	0m14.607s
user	3m55.179s

The 50s seems a touch excessive and it looks like a 20% parsing hit even when unused :(. Note the overall parsing times on this machine are fast as its running 24 in parallel. I'm going to try and figure out what is going on tomorrow as I've fixed a lot of this kind of thing in the past but wanted to mention it...

Cheers,

Richard

For those wanting to play:

bitbake core-image-sato -e -P

and looking at profile.log.processed shows 48 *million* calls to deepcopy. That is a big clue to the problem.

Cheers,

Richard

Oh, good thought.

Hmm. That doesn't totally surprise me.

The 50s seems... not entirely surprsing. I don't know that much about how Python does stuff, but in practice we end up with a LOT of deep copies of structures floating around because there are a ton of data copies; without these, finalize information gets duplicated and things go pretty wrong.

My intuition is that the stack backtrace thing is probably expensive, and might be noticably reduced by manually passing in the likely-missing values in cases where we know what they are, so it's less likely that any backtracing has to happen.

-s

Listen, get this.  Nobody with a good compiler needs to be justified.

Its not the stack backtrace stuff that is the limiting factor here, its the deepcopy, its really clear on the profiles. Anything with 48 million calls is a problem, we should be seeing 10s of thousands of calls tops for an operation like the above.

I tried a little experiment:

@@ -174,7 +174,14 @@ class IncludeHistory(object):
         self.filename = filename or '[TOP LEVEL]'
         self.children = []
         self.current = self
-        self.variables = {}
+        self.variables = COWDictBase.copy()
+
+    def copy(self):
+        new = IncludeHistory(self.parent, self.filename, self.datasmart)
+        new.children = copy.deepcopy(self.children)
+        new.current = self.current
+        new.variables = self.variables.copy()
+        return new
 
     def include(self, filename):
         newfile = IncludeHistory(self.current, filename)

@@ -615,7 +622,7 @@ class DataSmart(MutableMapping):
         # we really want this to be a DataSmart...
         data = DataSmart(seen=self._seen_overrides.copy(), special=self._special_values.copy())
         data.dict["_data"] = self.dict
-        data.history = copy.deepcopy(self.history)
+        data.history = self.history.copy()
         data.history.datasmart = data
         data._tracking = self._tracking
 

which replaces the brute force deepcopy with a more intelligent attempt to copy the IncludeHistory object. I've replaced the variables dict with our magic COW dict object since this is the kind of usage it was designed for.

What is interesting is that even with profiling enabled (-P), this drops the execution time to 10s from 50s so it proves my theory about deepcopy. Diffing the output before and after shows that it even faithfully handles the variable dependencies, it does however totally break the include list, as I break "current".

It does seem we're trying to merge two orthogonal sets of data into one object here (include history and variable history) and this results in a huge number of IncludeHistory instances in memory though the include chain. I'm therefore left wondering if we could separate the two?

Cheers,

Richard

I figured out the issue, I shouldn't have been touching current at all. Rather than deepcopy, its also better to use our own copy operation on the children:

@@ -174,7 +174,14 @@ class IncludeHistory(object):
         self.filename = filename or '[TOP LEVEL]'
         self.children = []
         self.current = self
-        self.variables = {}
+        self.variables = COWDictBase.copy()
+
+    def copy(self):
+        new = IncludeHistory(self.parent, self.filename, self.datasmart)
+        for c in self.children:
+            new.children.append(c.copy())
+        new.variables = self.variables.copy()
+        return new
 
     def include(self, filename):
         newfile = IncludeHistory(self.current, filename)

@@ -615,7 +622,7 @@ class DataSmart(MutableMapping):
         # we really want this to be a DataSmart...
         data = DataSmart(seen=self._seen_overrides.copy(), special=self._special_values.copy())
         data.dict["_data"] = self.dict
-        data.history = copy.deepcopy(self.history)
+        data.history = self.history.copy()
         data.history.datasmart = data
         data._tracking = self._tracking
 

So with this applied, we get -e taking around 10s and parsing 3m26 so about a 10% increase and the output of -e is the same before and after that patch. I'm still not happy but that is obviously a better proposition.

The real issue is these highly recursive and duplicate data structures are inefficient so I think there is still much room for improvement. Separating the variable and include history will likely help a lot.

The main point of the above is to prove how much the deepcopy hurts us and also that there are ways we can fix it.

Cheers,

Richard

To round this discussion out, I played with the code quite a bit and whilst I love the idea of the code, I found a few bugs and there were some structural issues in the way the code worked I didn't like. It was extremely easy to unintentionally break the history tracking.

This was due to many operations doing "fall though" with the set operation doing the final record, maybe with modified parameters.

Whatever we do, we need it to be robust enough to withstand other improvements in the codebase so this would have been enough to block it.

I decided to see if I could improve things so I've taken the patches and made quite some number of changes, I think the output with the revised version is either equal to or in some cases improved to the original. I've hopefully addresses the "robustness" issues with those changes as it feels less fragile to me and I removed some sources of confusion like the "value" verses "detail" business, standardising on "detail" in the end for a practicality reasons.

I still don't think the result is perfect however I do propose to merge it in the revised form I've just posted. I think any further optimisations or improvements can be done in tree.

Performance wise, with the code disabled, we're now down to 3m5 for overall parsing compared with 3m3 before which is in the region of statistical noise. The -e output is in around 10s which I think is reasonable for the functionality.

I appreciate its taken a while to get here with this. I'm hopeful the time waiting for review will be worth the revised patches.

Cheers,

Richard